mdw@git.distorted.org.uk Git - catacomb/blob - base/dispatch.c

   1 /* -*-c-*-
   2  *
   3  * CPU-specific dispatch
   4  *
   5  * (c) 2015 Straylight/Edgeware
   6  */
   7
   8 /*----- Licensing notice --------------------------------------------------*
   9  *
  10  * This file is part of Catacomb.
  11  *
  12  * Catacomb is free software; you can redistribute it and/or modify
  13  * it under the terms of the GNU Library General Public License as
  14  * published by the Free Software Foundation; either version 2 of the
  15  * License, or (at your option) any later version.
  16  *
  17  * Catacomb is distributed in the hope that it will be useful,
  18  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  19  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  20  * GNU Library General Public License for more details.
  21  *
  22  * You should have received a copy of the GNU Library General Public
  23  * License along with Catacomb; if not, write to the Free
  24  * Software Foundation, Inc., 59 Temple Place - Suite 330, Boston,
  25  * MA 02111-1307, USA.
  26  */
  27
  28 /*----- Header files ------------------------------------------------------*/
  29
  30 #include "config.h"
  31
  32 #include <ctype.h>
  33 #include <stdarg.h>
  34 #include <stdio.h>
  35 #include <stdlib.h>
  36 #include <string.h>
  37
  38 #include <mLib/macros.h>
  39
  40 #include "dispatch.h"
  41
  42 /*----- Intel x86/AMD64 feature probing -----------------------------------*/
  43
  44 #if CPUFAM_X86 || CPUFAM_AMD64
  45
  46 #  define CPUID1D_SSE2 (1u << 26)
  47 #  define CPUID1D_FXSR (1u << 24)
  48 #  define CPUID1C_PCLMUL (1u << 1)
  49 #  define CPUID1C_SSSE3 (1u << 9)
  50 #  define CPUID1C_AESNI (1u << 25)
  51 #  define CPUID1C_AVX (1u << 28)
  52 #  define CPUID1C_RDRAND (1u << 30)
  53
  54 struct cpuid { unsigned a, b, c, d; };
  55 extern int dispatch_x86ish_cpuid(struct cpuid *, unsigned a, unsigned c);
  56 extern int dispatch_x86ish_xmmregisters_p(void);
  57 extern int dispatch_x86ish_rdrand(unsigned *);
  58
  59 static void cpuid(struct cpuid *cc, unsigned a, unsigned c)
  60 {
  61   int rc = dispatch_x86ish_cpuid(cc, a, c);
  62   if (rc)
  63     dispatch_debug("CPUID instruction not available");
  64   else
  65     dispatch_debug("CPUID(%08x, %08x) -> %08x, %08x, %08x, %08x",
  66                    a, c, cc->a, cc->b, cc->c, cc->d);
  67 }
  68
  69 static unsigned cpuid_maxleaf(void)
  70   { struct cpuid c; cpuid(&c, 0, 0); return (c.a); }
  71
  72 /* --- @cpuid_features_p@ --- *
  73  *
  74  * Arguments:   @unsigned dbits@ = bits to check in EDX
  75  *              @unsigned cbits@ = bits to check in ECX
  76  *
  77  * Returns:     Nonzero if all the requested bits are set in the CPUID result
  78  *              on leaf 1.
  79  */
  80
  81 static int cpuid_features_p(unsigned dbits, unsigned cbits)
  82 {
  83   struct cpuid c;
  84   if (cpuid_maxleaf() < 1) return (0);
  85   cpuid(&c, 1, 0);
  86   return ((c.d & dbits) == dbits && (c.c & cbits) == cbits);
  87 }
  88
  89 /* --- @xmm_registers_available_p@ --- *
  90  *
  91  * Arguments:   ---
  92  *
  93  * Returns:     Nonzero if the operating system has made the XMM registers
  94  *              available for use.
  95  */
  96
  97 static int xmm_registers_available_p(void)
  98 {
  99   int f = dispatch_x86ish_xmmregisters_p();
 100
 101   dispatch_debug("XMM registers %savailable", f ? "" : "not ");
 102   return (f);
 103 }
 104
 105 /* --- @rdrand_works_p@ --- *
 106  *
 107  *
 108  * Arguments:   ---
 109  *
 110  * Returns:     Nonzero if the `rdrand' instruction actually works.  Assumes
 111  *              that it's already been verified to be safe to issue.
 112  */
 113
 114 static int rdrand_works_p(void)
 115 {
 116   unsigned ref, x, i;
 117
 118   /* Check that it doesn't always give the same answer.  Try four times: this
 119    * will fail with probability %$2^{-128}$% with a truly random generator,
 120    * which seems fair enough.
 121    */
 122   if (dispatch_x86ish_rdrand(&ref)) goto fail;
 123   for (i = 0; i < 4; i++) {
 124     if (dispatch_x86ish_rdrand(&x)) goto fail;
 125     if (x != ref) goto not_stuck;
 126   }
 127   dispatch_debug("RDRAND always returns 0x%08x!", ref);
 128   return (0);
 129
 130 not_stuck:
 131   dispatch_debug("RDRAND instruction looks plausible");
 132   return (1);
 133
 134 fail:
 135   dispatch_debug("RDRAND instruction fails too often");
 136   return (0);
 137 }
 138
 139 #endif
 140
 141 /*----- General feature probing using auxiliary vectors -------------------*/
 142
 143 /* Try to find the system's definitions for auxiliary vector entries. */
 144 #ifdef HAVE_SYS_AUXV_H
 145 #  include <sys/auxv.h>
 146 #endif
 147 #ifdef HAVE_LINUX_AUXVEC_H
 148 #  include <linux/auxvec.h>
 149 #endif
 150 #ifdef HAVE_ASM_HWCAP_H
 151 #  include <asm/hwcap.h>
 152 #endif
 153
 154 /* The type of entries in the auxiliary vector.  I'm assuming that `unsigned
 155  * long' matches each platform's word length; if this is false then we'll
 156  * need some host-specific tweaking here.
 157  */
 158 union auxval { long i; unsigned long u; const void *p; };
 159 struct auxentry { unsigned long type; union auxval value; };
 160
 161 /* Register each CPU family's interest in the auxiliary vector.  Make sure
 162  * that the necessary entry types are defined.  This is primarily ordered by
 163  * entry type to minimize duplication.
 164  */
 165 #if defined(AT_HWCAP) && CPUFAM_ARMEL
 166 #  define WANT_ANY 1
 167 #  define WANT_AT_HWCAP(_) _(AT_HWCAP, u, hwcap)
 168 #endif
 169
 170 #if defined(AT_HWCAP) && CPUFAM_ARM64
 171 #  define WANT_ANY 1
 172 #  define WANT_AT_HWCAP(_) _(AT_HWCAP, u, hwcap)
 173 #endif
 174
 175 #if defined(AT_HWCAP2) && CPUFAM_ARMEL
 176 #  define WANT_ANY 1
 177 #  define WANT_AT_HWCAP2(_) _(AT_HWCAP2, u, hwcap2)
 178 #endif
 179
 180 /* If we couldn't find any interesting entries then we can switch all of this
 181  * machinery off.  Also do that if we have no means for atomic updates.
 182  */
 183 #if WANT_ANY && CPU_DISPATCH_P
 184
 185 /* The main output of this section is a bitmask of detected features.  The
 186  * least significant bit will be set if we've tried to probe.  Always access
 187  * this using `DISPATCH_LOAD' and `DISPATCH_STORE'.
 188  */
 189 static unsigned hwcaps = 0;
 190
 191 /* For each potentially interesting type which turned out not to exist or be
 192  * wanted, define a dummy macro for the sake of the next step.
 193  */
 194 #ifndef WANT_AT_HWCAP
 195 #  define WANT_AT_HWCAP(_)
 196 #endif
 197 #ifndef WANT_AT_HWCAP2
 198 #  define WANT_AT_HWCAP2(_)
 199 #endif
 200
 201 /* For each CPU family, define two lists.
 202  *
 203  *   * `WANTAUX' is a list of the `WANT_AT_MUMBLE' macros which the CPU
 204  *     family tried to register interest in above.  Each entry contains the
 205  *     interesting auxiliary vector entry type, the name of the union branch
 206  *     for its value, and the name of the slot in `struct auxprobe' in which
 207  *     to store the value.
 208  *
 209  *   * `CAPMAP' is a list describing the output features which the CPU family
 210  *     intends to satisfy from the auxiliary vector.  Each entry contains a
 211  *     feature name suffix, and the token name (for `check_env').
 212  */
 213 #if CPUFAM_ARMEL
 214 #  define WANTAUX(_)                                                    \
 215         WANT_AT_HWCAP(_)                                                \
 216         WANT_AT_HWCAP2(_)
 217 #  define CAPMAP(_)                                                     \
 218         _(ARM_VFP, "arm:vfp")                                           \
 219         _(ARM_NEON, "arm:neon")                                         \
 220         _(ARM_V4, "arm:v4")                                             \
 221         _(ARM_D32, "arm:d32")                                           \
 222         _(ARM_AES, "arm:aes")                                           \
 223         _(ARM_PMULL, "arm:pmull")
 224 #endif
 225 #if CPUFAM_ARM64
 226 #  define WANTAUX(_)                                                    \
 227         WANT_AT_HWCAP(_)
 228 #  define CAPMAP(_)                                                     \
 229         _(ARM_AES, "arm:aes")                                           \
 230         _(ARM_PMULL, "arm:pmull")
 231 #endif
 232
 233 /* Build the bitmask for `hwcaps' from the `CAPMAP' list. */
 234 enum {
 235   HFI_PROBED = 0,
 236 #define HFI__ENUM(feat, tok) HFI_##feat,
 237   CAPMAP(HFI__ENUM)
 238 #undef HFI__ENUM
 239   HFI__END
 240 };
 241 enum {
 242   HF_PROBED = 1,
 243 #define HF__FLAG(feat, tok) HF_##feat = 1 << HFI_##feat,
 244   CAPMAP(HF__FLAG)
 245 #undef HF__FLAG
 246   HF__END
 247 };
 248
 249 /* Build a structure in which we can capture the interesting data from the
 250  * auxiliary vector.
 251  */
 252 #define AUXUTYPE_i long
 253 #define AUXUTYPE_u unsigned long
 254 #define AUXUTYPE_p const void *
 255 struct auxprobe {
 256 #define AUXPROBE__SLOT(type, ubranch, slot) AUXUTYPE_##ubranch slot;
 257   WANTAUX(AUXPROBE__SLOT)
 258 #undef AUXPROBE_SLOT
 259 };
 260
 261 /* --- @probe_hwcaps@ --- *
 262  *
 263  * Arguments:   ---
 264  *
 265  * Returns:     ---
 266  *
 267  * Use:         Attempt to find the auxiliary vector (which is well hidden)
 268  *              and discover interesting features from it.
 269  */
 270
 271 static void probe_hwcaps(void)
 272 {
 273   unsigned hw = HF_PROBED;
 274   struct auxprobe probed = { 0 };
 275
 276   /* Populate `probed' with the information we manage to retrieve from the
 277    * auxiliary vector.  Slots we couldn't find are left zero-valued.
 278    */
 279 #if defined(HAVE_GETAUXVAL)
 280   /* Shiny new libc lets us request individual entry types.  This is almost
 281    * too easy.
 282    */
 283 #  define CAP__GET(type, ubranch, slot)                                 \
 284         probed.slot = (AUXUTYPE_##ubranch)getauxval(type);
 285   WANTAUX(CAP__GET)
 286 #else
 287   /* Otherwise we're a bit stuck, really.  Modern Linux kernels make a copy
 288    * of the vector available in `/procc' so we could try that.
 289    *
 290    * The usual place is stuck on the end of the environment vector, but that
 291    * may well have moved, and we have no way of telling whether it has or
 292    * whether there was ever an auxiliary vector there at all; so don't do
 293    * that.
 294    */
 295   {
 296     FILE *fp = 0;
 297     unsigned char *p = 0, *q = 0;
 298     const struct auxentry *a;
 299     size_t sz, off, n;
 300
 301     /* Open the file and read it into a memory chunk. */
 302     if ((fp = fopen("/proc/self/auxv", "rb")) == 0) goto clean;
 303     sz = 4096; off = 0;
 304     if ((p = malloc(sz)) == 0) goto clean;
 305     for (;;) {
 306       n = fread(p + off, 1, sz - off, fp);
 307       off += n;
 308       if (off < sz) break;
 309       sz *= 2; if ((q = realloc(p, sz)) == 0) break;
 310       p = q;
 311     }
 312
 313     /* Work through the vector (or as much of it as we found) and extract the
 314      * types we're interested in.
 315      */
 316     for (a = (const struct auxentry *)p,
 317            n = sz/sizeof(struct auxentry);
 318          n--; a++) {
 319       switch (a->type) {
 320 #define CAP__SWITCH(type, ubranch, slot)                                \
 321         case type: probed.slot = a->value.ubranch; break;
 322         WANTAUX(CAP__SWITCH)
 323         case AT_NULL: goto clean;
 324       }
 325     }
 326
 327   clean:
 328     if (p) free(p);
 329     if (fp) fclose(fp);
 330   }
 331 #endif
 332
 333   /* Each CPU family now has to pick through what was found and stashed in
 334    * `probed', and set the appropriate flag bits in `hw'.
 335    */
 336 #if CPUFAM_ARMEL
 337   if (probed.hwcap & HWCAP_VFPv3) hw |= HF_ARM_VFP;
 338   if (probed.hwcap & HWCAP_NEON) hw |= HF_ARM_NEON;
 339   if (probed.hwcap & HWCAP_VFPD32) hw |= HF_ARM_D32;
 340   if (probed.hwcap & HWCAP_VFPv4) hw |= HF_ARM_V4;
 341 #  ifdef HWCAP2_AES
 342   if (probed.hwcap2 & HWCAP2_AES) hw |= HF_ARM_AES;
 343 #  endif
 344 #  ifdef HWCAP2_PMULL
 345   if (probed.hwcap2 & HWCAP2_PMULL) hw |= HF_ARM_PMULL;
 346 #  endif
 347 #endif
 348 #if CPUFAM_ARM64
 349   if (probed.hwcap & HWCAP_AES) hw |= HF_ARM_AES;
 350   if (probed.hwcap & HWCAP_PMULL) hw |= HF_ARM_PMULL;
 351 #endif
 352
 353   /* Store the bitmask of features we probed for everyone to see. */
 354   DISPATCH_STORE(hwcaps, hw);
 355
 356   /* Finally, make a report about the things we found.  (Doing this earlier
 357    * will pointlessly widen the window in which multiple threads will do the
 358    * above auxiliary-vector probing.)
 359    */
 360 #define CAP__DEBUG(feat, tok)                                           \
 361   dispatch_debug("check auxv for feature `%s': %s", tok,                \
 362                  hw & HF_##feat ? "available" : "absent");
 363   CAPMAP(CAP__DEBUG)
 364 #undef CAP__DEBUG
 365 }
 366
 367 /* --- @get_hwcaps@ --- *
 368  *
 369  * Arguments:   ---
 370  *
 371  * Returns:     A mask of hardware capabilities and other features, as probed
 372  *              from the auxiliary vector.
 373  */
 374
 375 static unsigned get_hwcaps(void)
 376 {
 377   unsigned hw;
 378
 379   DISPATCH_LOAD(hwcaps, hw);
 380   if (!(hwcaps & HF_PROBED)) { probe_hwcaps(); DISPATCH_LOAD(hwcaps, hw); }
 381   return (hw);
 382 }
 383
 384 #endif
 385
 386 /*----- External interface ------------------------------------------------*/
 387
 388 /* --- @dispatch_debug@ --- *
 389  *
 390  * Arguments:   @const char *fmt@ = a format string
 391  *              @...@ = additional arguments
 392  *
 393  * Returns:     ---
 394  *
 395  * Use:         Writes a formatted message to standard output if dispatch
 396  *              debugging is enabled.
 397  */
 398
 399 void dispatch_debug(const char *fmt, ...)
 400 {
 401   va_list ap;
 402   const char *e = getenv("CATACOMB_CPUDISPATCH_DEBUG");
 403
 404   if (e && *e != 'n' && *e != '0') {
 405     va_start(ap, fmt);
 406     fputs("Catacomb CPUDISPATCH: ", stderr);
 407     vfprintf(stderr, fmt, ap);
 408     fputc('\n', stderr);
 409     va_end(ap);
 410   }
 411 }
 412
 413 /* --- @check_env@ --- *
 414  *
 415  * Arguments:   @const char *ftok@ = feature token
 416  *
 417  * Returns:     Zero if the feature is forced off; positive if it's forced
 418  *              on; negative if the user hasn't decided.
 419  *
 420  * Use:         Checks the environment variable `CATACOMB_CPUFEAT' for the
 421  *              feature token @ftok@.  The variable, if it exists, should be
 422  *              a space-separated sequence of `+tok' and `-tok' items.  These
 423  *              tokens may end in `*', which matches any suffix.
 424  */
 425
 426 static int IGNORABLE check_env(const char *ftok)
 427 {
 428   const char *p, *q, *pp;
 429   int d;
 430
 431   p = getenv("CATACOMB_CPUFEAT");
 432   if (!p) return (-1);
 433
 434   for (;;) {
 435     while (isspace((unsigned char)*p)) p++;
 436     if (!*p) return (-1);
 437     switch (*p) {
 438       case '+': d = +1; p++; break;
 439       case '-': d =  0; p++; break;
 440       default:  d = -1;      break;
 441     }
 442     for (q = p; *q && !isspace((unsigned char)*q); q++);
 443     if (d >= 0) {
 444       for (pp = ftok; p < q && *pp && *p == *pp; p++, pp++);
 445       if ((p == q && !*pp) || (*p == '*' && p + 1 == q)) return (d);
 446     }
 447     p = q;
 448   }
 449   return (-1);
 450 }
 451
 452 /* --- @cpu_feature_p@ --- *
 453  *
 454  * Arguments:   @unsigned feat@ = a @CPUFEAT_...@ code
 455  *
 456  * Returns:     Nonzero if the feature is available.
 457  */
 458
 459 #include <stdio.h>
 460
 461 static int IGNORABLE
 462   feat_debug(const char *ftok, const char *check, int verdict)
 463 {
 464   if (verdict >= 0) {
 465     dispatch_debug("feature `%s': %s -> %s", ftok, check,
 466                    verdict ? "available" : "absent");
 467   }
 468   return (verdict);
 469 }
 470
 471 int cpu_feature_p(int feat)
 472 {
 473   int IGNORABLE f;
 474   IGNORE(f);
 475 #define CASE_CPUFEAT(feat, ftok, cond) case CPUFEAT_##feat:             \
 476   if ((f = feat_debug(ftok, "environment override",                     \
 477                       check_env(ftok))) >= 0)                           \
 478     return (f);                                                         \
 479   else                                                                  \
 480     return (feat_debug(ftok, "runtime probe", cond));
 481
 482   switch (feat) {
 483 #if CPUFAM_X86 || CPUFAM_AMD64
 484     CASE_CPUFEAT(X86_SSE2, "x86:sse2",
 485                  cpuid_features_p(CPUID1D_SSE2, 0) &&
 486                  xmm_registers_available_p());
 487     CASE_CPUFEAT(X86_AESNI, "x86:aesni",
 488                  cpuid_features_p(CPUID1D_SSE2, CPUID1C_AESNI) &&
 489                  xmm_registers_available_p());
 490     CASE_CPUFEAT(X86_RDRAND, "x86:rdrand",
 491                  cpuid_features_p(0, CPUID1C_RDRAND) && rdrand_works_p());
 492     CASE_CPUFEAT(X86_AVX, "x86:avx",
 493                  cpuid_features_p(0, CPUID1C_AVX) &&
 494                  xmm_registers_available_p());
 495     CASE_CPUFEAT(X86_SSSE3, "x86:ssse3",
 496                  cpuid_features_p(0, CPUID1C_SSSE3) &&
 497                  xmm_registers_available_p());
 498     CASE_CPUFEAT(X86_PCLMUL, "x86:pclmul",
 499                  cpuid_features_p(0, CPUID1C_PCLMUL) &&
 500                  xmm_registers_available_p());
 501 #endif
 502 #ifdef CAPMAP
 503 #  define FEATP__CASE(feat, tok)                                        \
 504         CASE_CPUFEAT(feat, tok, get_hwcaps() & HF_##feat)
 505     CAPMAP(FEATP__CASE)
 506 #undef FEATP__CASE
 507 #endif
 508     default:
 509       dispatch_debug("denying unknown feature %d", feat);
 510       return (0);
 511   }
 512 #undef CASE_CPUFEAT
 513 }
 514
 515 /*----- That's all, folks -------------------------------------------------*/